[RFA,1/2] Speed up dict_hash
Commit Message
This speeds up dict_hash a bit, by moving the "TKB" check into the
switch in the loop.
For "gdb -nx -readnow -batch gdb", this improves the time from ~9.8s
before to ~8.5s afterward.
gdb/ChangeLog
2017-11-04 Tom Tromey <tom@tromey.com>
* dictionary.c (dict_hash): Move "TKB" check into the "switch".
---
gdb/ChangeLog | 4 ++++
gdb/dictionary.c | 32 ++++++++++++++++----------------
2 files changed, 20 insertions(+), 16 deletions(-)
Comments
On 11/04/2017 04:13 PM, Tom Tromey wrote:
> This speeds up dict_hash a bit, by moving the "TKB" check into the
> switch in the loop.
>
> For "gdb -nx -readnow -batch gdb", this improves the time from ~9.8s
> before to ~8.5s afterward.
Nice!
> + case 'T':
> + /* Ignore "TKB" suffixes.
> +
> + These are used by Ada for subprograms implementing a task body.
> + For instance for a task T inside package Pck, the name of the
> + subprogram implementing T's body is `pck__tTKB'. We need to
> + ignore the "TKB" suffix because searches for this task body
> + subprogram are going to be performed using `pck__t' (the encoded
> + version of the natural name `pck.t'). */
> + if (strcmp (string, "TKB") == 0)
> + return hash;
This looks good to me.
OOC, did you check whether
'if (strcmp (string + 1, "KB") == 0)'
or even:
if (string[1] == 'K' && string[2] == 'B' && string[3] == '\0')
made a difference?
Maybe there aren't enough 'T's in symbols to matter, or the
compiler is already inlining that strcmp...
Thanks,
Pedro Alves
>>>>> "Pedro" == Pedro Alves <palves@redhat.com> writes:
>> For "gdb -nx -readnow -batch gdb", this improves the time from ~9.8s
>> before to ~8.5s afterward.
Pedro> Nice!
I rebuilt today and the baseline times were higher -- now it starts at
~14.1s for an unmodified gdb, down to ~12.2 with my patch. I'm not sure
if this is due to some other patch, or if the additional time is because
I configured differently (I enabled guile, python, and all targets in
this most recent build). If so, 4 seconds (comparing to the ~10 seconds
last time around) seems awfully expensive for those features...
Pedro> OOC, did you check whether
Pedro> 'if (strcmp (string + 1, "KB") == 0)'
Pedro> or even:
Pedro> if (string[1] == 'K' && string[2] == 'B' && string[3] == '\0')
Pedro> made a difference?
I tried these and they didn't make a difference.
Tom
@@ -796,17 +796,6 @@ dict_hash (const char *string0)
hash = 0;
while (*string)
{
- /* Ignore "TKB" suffixes.
-
- These are used by Ada for subprograms implementing a task body.
- For instance for a task T inside package Pck, the name of the
- subprogram implementing T's body is `pck__tTKB'. We need to
- ignore the "TKB" suffix because searches for this task body
- subprogram are going to be performed using `pck__t' (the encoded
- version of the natural name `pck.t'). */
- if (strcmp (string, "TKB") == 0)
- return hash;
-
switch (*string)
{
case '$':
@@ -828,14 +817,25 @@ dict_hash (const char *string0)
return hash;
hash = 0;
string += 2;
- break;
+ continue;
}
- /* FALL THROUGH */
- default:
- hash = SYMBOL_HASH_NEXT (hash, *string);
- string += 1;
+ break;
+ case 'T':
+ /* Ignore "TKB" suffixes.
+
+ These are used by Ada for subprograms implementing a task body.
+ For instance for a task T inside package Pck, the name of the
+ subprogram implementing T's body is `pck__tTKB'. We need to
+ ignore the "TKB" suffix because searches for this task body
+ subprogram are going to be performed using `pck__t' (the encoded
+ version of the natural name `pck.t'). */
+ if (strcmp (string, "TKB") == 0)
+ return hash;
break;
}
+
+ hash = SYMBOL_HASH_NEXT (hash, *string);
+ string += 1;
}
return hash;
}